Table of Contents

RFC 0013 — The Narrative Obligation Ledger

Status: approved by the author 2026-09-15; Phases 0–2 implemented the same day (mojibake convergence + ingress guards; ledger schema, write-door extraction, Brief block, trial balance, author desk; reconciliation instrument, resurfacing judge, record grounding, calibration harness, health snapshots, publish gate 6, --auto-run hard gate). Awaits the author's Hub redeploy; the two migrations apply on its next start. Phase 3 (BCODA retrofit) runs after calibration on the gutenberg fixtures meets the bar in §6. · Companions: RFC 0009 (no autonomous prose writes), RFC 0012 (single-source writer), docs/LOGIC.md §3.4 and §9, docs/LEDGER.md, docs/READER-QA.md §6.

0. The law

A narrative promise is logged as an obligation at the moment it is made, with a verbatim quote and the beat it came from. An obligation is closed only by a quote from the beat that pays it. An obligation that is open past its due point without an author decision is a finding. A payoff with no origin is a finding. An empty ledger fails, it does not pass. Nothing in this system writes prose.

1. Why

Reading the first five pages of Bushido Coda found three defects no instrument had ever flagged:

  1. A girl watching from behind a curtain in Ch1 — half a page of weighted setup — never mentioned again in ~475 beats (verified by full-text search of the whole book).
  2. A second unnamed child hidden in loading-dock ductwork through a six-on-one fight, with no on-page reason for a child to be at a professional ambush.
  3. Mrs. Chen's character record carrying a fabricated backstory ("mercy job… sickbed") that exists in no beat.

The book's 140k-character node bible is a real, maintained plan and mentions none of them: they were invented at prose-generation time and never fed back. The root cause, located in code:

  • the writer sees only the last 6,000 characters of the current chapter (BeatGeneratorService.cs, context.SceneSoFar[^6000..]); earlier chapters arrive as one-line summaries;
  • the old NodeOpenThreads extraction ran only after Generation — BCODA was hand-spliced, so its promises were never recorded; the rows it did make had no quote, no due point, no entity, no read path;
  • PlantPayoffs for BCODA was empty: the bible's §14 ledger existed only as markdown;
  • EntityMentionScanner tags names only; a lowercase referent ("the girl") has no record and no stub; FactInterpreterService.CreateStubAsync was never wired to a beat write;
  • docs/READER-QA.md §6 classified an unresolved mystery as a non-finding.

The author's question was the right one: no model (and no editor) holds a 500-page novel in working memory. Lost in the Middle (TACL 2024), NoCha (EMNLP 2024: GPT-4o 55.8% on whole-novel claims), and TLDM (2025: storyworld tracking collapses past ~64k tokens) show recall does not scale. Re3/DOC (+14% / +22.5% coherence from plan-controlled generation), CFPG (2026: Foreshadow–Trigger–Payoff pools; payoff realization 0.569 → 0.911 with oracle timing) and Lost in Stories (2026: a quote-grounded LLM checker validated on 1,000 injected errors reaches F1 0.678 against professional annotators' 0.281) show what does: record the promise when it is made, verify closure against the text, calibrate the instrument on injected defects. Every mature fact-keeping discipline — double-entry bookkeeping, DO-178C bidirectional traceability, legal fact chronologies, script-supervisor continuity logs, copyeditor style sheets — converged on the same shape: an obligation opened at introduction, carrying provenance, explicitly closed, listable while open, verified against the artifact.

2. The ledger (Phase 1, shipped)

NarrativeObligations — book-scoped rows: Kind (promise | plant | question | wound | foreshadow | introduced-referent | unexplained-presence), Description, Provenance (ClaimProvenance: authored | observed | inferred), OriginBeatId + OriginQuote + OriginTextHash, optional EntityId (a canon entity or a Status="stub" row named (unnamed) girl behind the curtain — the ( keeps EntityMentionScanner off it), CFPG TriggerCondition, DueByKind/DueByValue (chapter | beats | book-end), State (Open → Advanced → Closed | Dropped | Deferred, plus system-only Withdrawn), closing beat + quote + hash, AuthorNote, DroppedReason, AuthorLocked, DedupKey. NarrativeObligationEvents is the journal (open / advance / close / reopen / drop / defer / withdraw / reanchor / due-changed / lock, with actor). PlantPayoffs.ObligationId bridges hand-curated pairs; Beats.ObligationScanHash gates the scan. Migration 20260915185326_AddNarrativeObligations moved every NodeOpenThreads row across as inferred, mirrored every plant pair as an authored, locked plant row, then dropped the old table.

The door. NarrativeObligationService.ScanBeatAsync is hooked in NodeWorkbenchService.UpdateBeatTextAsync (and the batch path) for every BeatWriteReason except TagMaintenance/Plan, fire-and-forget, hash-gated on whitespace-collapsed text. It first re-verifies the quotes this beat anchors — restamp, re-anchor to a sibling beat, reopen a closure whose quote vanished, withdraw an open unlocked row nothing ever advanced — then one Haiku call (NarrativeObligationExtractor, reason-before-verdict JSON) returns what the beat opened and which open rows it advanced or closed. Every quote is checked in code (Audit.QuoteGrounding, shared with the logic sweep); an item that cannot quote the text is discarded and counted. A provider outage leaves the gate unstamped so the beat is re-scanned next time. Locked rows are immutable to every automated path.

The Brief. BuildBriefBlockAsync replaces the OPEN THREADS block and the separate PLANTED DETAILS block: [OPEN OBLIGATIONS] with DUE NOW / NOT YET, urgency-ordered, ≤2,500 chars.

The desk. prose --obligations list|trial-balance|history|open|close|drop|defer|reopen|due| accept|rescan and MCP list_obligations, get_obligation, obligation_trial_balance, open_obligation, close_obligation (refused with quote_not_found unless the quote is in that beat), drop_obligation (reason + note required), defer_obligation, reopen_obligation, set_obligation_due, accept_obligation, link_obligation_entity. Every author write locks the row and journals the actor. "I'll pay it in Ch12" is a Defer to Ch12 — no state, no code path, could write the payoff.

The bible's own ledger. prose --obligations import-bible-ledger --slug S [--dry-run] (BibleLedgerImporter) reads the node bible's §14a "Closed plants" and §14b "Dropped findings" tables from NodeOutlineSections (or the legacy Nodes.NodeOutline) and lands them as authored, locked rows: §14a → plant, Closed, mirrored into PlantPayoffs; §14b → Dropped with DroppedReason = author-note and the bible's reason verbatim. Anchors resolve only when a Ch<n> SK:<k> label matches a beat's SortKey exactly in the current reading order; anything else is imported with a null anchor and printed first as NEEDS ANCHOR — a lossy label is never guessed. Rows carry no page quote (the bible is the author's word), so the trial balance treats them as neither stale nor dangling. Idempotent via DedupKey; no §14 table at all is COULD NOT LOOK, exit 1.

3. The trial balance and the hard gate (author decision: hard)

TrialBalanceAsync(book, chapter): opened − closed − dropped − deferred = carried forward, plus overdue without decision — outstanding rows past due at the end of the chapter, not locked into an author decision. Wired in: prose --auto-run --obligation-gate hard|soft|off (default hard: scans the chapter's beats synchronously, strikes the balance, stops the run with exit 3 when a chapter owes an undecided debt); publish_readiness check 6 (LOGIC.md §9 item 6 — scan coverage 100%, zero overdue undecided, zero open book-end rows at publish; an empty ledger FAILS).

4. Instruments (Phase 2, shipped)

Command Cost What
prose --reconcile-obligations --slug S [--json] / --all [--status …] free six deterministic rules via AuditRunner under FindingCategory.NarrativeObligation, scope node:{slug}#obligations: overdue_open (BLOCKER for an early unnamed referent/question never advanced), open_at_end, stale_closure, dangling_beat, unplanted_payoff (LOGIC §3.4 reverse direction), deferred_expired. Writes a NarrativeHealthSnapshots row when it examined anything. Prints examined N over M beats; COULD NOT LOOK on an empty ledger.
… --deep cents (own cost name --reconcile-obligations-deep) ObligationResurfacingJudge: per open row, candidates from embeddings (FindSimilarBeatNodesAsync) ∪ lexical overlap, one Haiku call, verdicts closes/advances/not_addressed with verbatim quote, ungrounded discarded, cached in ObligationJudgeCache by candidate text hash.
prose --scan-unnamed-referents --slug S free UnnamedReferentScanner: definite NPs with a person/role head and no tag; flags ≥2 mentions in one beat, acting, never seen again. Hint-only.
prose --ground-entity-records --slug S [--entity N] cents EntityRecordGroundingService: decompose each tagged entity's record into atomic claims (FActScore); entail against tagged + embedding-retrieved beats with a quote gate; unentailed/contradicted filed under EntityDrift (node:{slug}#recordground); matching non-authored ContinuityClaims downgraded to inferred. Never edits the record.
prose --inject-calibration-defects / --calibrate-obligations / --revert-calibration-defects cents §6. gutenberg universe only; throws before any write elsewhere.
prose --narrative-health --slug S [--history] free the snapshots.

5. Deliberate vs abandoned

Decided by a ledger row, never by prose language: a Deferred row, or an authored row due at book-end, is not flagged before the end. READER-QA §6's demotion stays; what it demotes should now become a question row so the trial balance can age it instead of dropping it.

6. Calibration bar (Phase 2 → 3 gate)

On GCNEG / GCSH / GCTOC with N = 8–20 injected defects (half abandoned, half resolved): precision ≥ 0.85, recall ≥ 0.70, F1 ≥ 0.678 (the Lost in Stories bar), ≤ 1.0 MODERATE+ control finding per 10k words on the un-injected text, ≤ 10% of resolved injections mis-flagged. Any extractor or judge prompt change re-runs the harness before it ships (the battery value audit (BIBLE ADR-10): re-admission by evidence).

6a. GCTOC answer key — what Book the First actually owes (hand-read 2026-09-16)

Written by reading the whole fixture before looking at any extracted ledger, so the instrument is graded against the book rather than against a rationalisation of its own output.

Opened AND paid inside Book the First — a competent ledger must open and close each of these:

# debt origin quote paid by
1 What "RECALLED TO LIFE" means Jerry, say that my answer was, RECALLED TO LIFE. he has been--been found. He is alive. (Ch4)
2 Who "Mam'selle" is Wait at Dover for Mam'selle. Miss Manette had arrived from London (Ch4)
3 Who the buried man of Lorry's dream is Buried how long? / Almost eighteen years. Dr Manette, 18 years in the North Tower (Ch4–6)
4 Whether Lucie's father is really dead my poor father, whom I never saw--so long dead Your father has been taken to the house of an old servant in Paris (Ch4)
5 Who the white-haired shoemaker is a white-haired man sat on a low bench, stooping forward and very busy, making shoes Monsieur Manette, do you remember nothing of me? (Ch6)
6 What the golden hairs in the folded rag are it contained a very little quantity of hair She had laid her head upon my shoulder, that night when I was summoned out (Ch6)

Legitimately CARRIED past the end of Book the First — Dickens pays these in Books Two and Three. A correct ledger must hold them OPEN, and must not be scored for doing so:

# debt origin quote paid in
7 Why Manette was imprisoned, and by whom worse than useless to seek to know whether he has been for years overlooked, or always designedly held prisoner Book the Third (the letter)
8 Jerry Cruncher's night trade Jerry, you honest tradesman, it wouldn't suit _your_ line of business! Book the Second (Resurrection-Man)
9 The wine-cask / BLOOD foreshadow scrawled upon a wall with his finger dipped in muddy wine-lees--BLOOD Book the Third
10 The Woodman and the Farmer to make a certain movable framework with a sack and a knife in it, terrible in history Book the Third (guillotine, tumbrils)
11 The lamplighter's ropes and pulleys hauling up men by those ropes and pulleys Book the Second (Foulon)
12 Madame Defarge's knitting took up her knitting with great apparent calmness and repose of spirit, and became absorbed in it Books Two/Three (the register)
13 "Jacques" as a shared name I choose them as real men, of my name--Jacques is my name Book the Second (the Jacquerie)
14 The unnamed red-haired woman a wild-looking woman … all of a red colour Book the Second (Miss Pross)
15 Manette's relapse risk he would be frightened--rave--tear himself to pieces--die Book the Second (after the wedding)

The consequence for the bar, and it is not about the instrument. A perfect reader of this fixture ends with ≈9 obligations outstanding across 17,300 words. The calibration bar allows ≤ 1.0 MODERATE+ control finding per 10k words ≈ 1.7 for the whole book. So a flawless ledger on A Tale of Two Cities, Book the First scores roughly five times the permitted control rate — by being right. The control metric treats outstanding-at-end as a false positive, which holds for GCSH (twelve self-contained stories that close what they open) and is simply false for the first act of a three-book novel, where unpaid debts are the structure. BCODA is a novel, not a story collection. Before any further tuning is bought, the bar needs a rule that distinguishes a debt the text abandons from a debt the text has not reached yet — most likely by scoring control findings only against a text that is structurally complete, or by scoring "outstanding at end" separately from "contradicted / abandoned". Until then, every precision number in this log is measured against a target that penalises correct behaviour on exactly the shape of book the engine exists to write.

7. Out of scope, by law

Automatic payoff generation; rewriting accepted prose; a "fix all orphans" command; LLM ranking of obligations by importance; per-universe rulebooks. Findings describe, never instruct.

8. Status log

  • 2026-09-15 — Phase 0 93c7cdb74; Phase 1 fae87dc55; Phase 2 this commit. Hub redeploy and calibration run are the author's next steps; then the BCODA retrofit runbook (plan file mellow-sauteeing-squirrel.md, "BCODA retrofit runbook").

  • 2026-09-15 (later) — Hub redeployed; BCODA bible mojibake CLEAN in one pass. The gutenberg fixtures were one beat per chapter (12/6/3) and the harness needs ≥20, so GCSH was split to 96 beats and GCTOC to 17 at sentence boundaries (deterministic split_beat, verbatim); both marked "Complete - publication ready" so book-end rules age. Calibration run 1 on GCSH: BELOW BAR. --deep died after $2.02 (est $0.05) on ObligationJudgeCache.Quote nvarchar(400) overflow; the free-rules score was TP 1 / FN 3 / resolved mis-flagged 2 / 116 control MODERATE+ (11.1 per 10k words), P 0.008 R 0.25. Two root causes found by reading, both fixed in this commit: (1) the extractor cut every beat at 6,000 chars before the prompt AND the quote gate, so the tail of any longer beat — where injection appends — could never be opened or recognised as a payoff; v2 scans long beats in sentence-boundary windows gated against the whole beat, and PromptVersion is now folded into ObligationScanHash so the rescan is not skipped; (2) every persisted quote is now QuoteGrounding.ClampForStorage (≤400, word boundary; a prefix of a substring still grounds). The 116 control findings are open obligations the free rules cannot close — the deep judge is the instrument that closes them, and it never ran; that number is unmeasured, not failed. Measured beat lengths (raw chars) attribute the misses precisely: the lantern boy (6,597) and stopped clock (6,648) setups and the brass-whistle (6,603) and tan-coat (6,825) payoffs sat past the cut; so did the one-eyed-cat setup (6,347), which the score counted "ok — never opened" and was in fact a hidden miss. Two failures were NOT truncation: the violet-seal setup (5,525 chars, fully visible; that beat logged 2 ungrounded quotes — quote fidelity, e.g. "recognised" respelled) and the tin-soldier payoff (5,815, visible, not recognised as closing — suspect MaxOpenListed = 40 saturating the listed open set on a 116-open book, or an "advanced" verdict). Run 2 decides whether those two need extractor work; the windowing alone cannot fix them. Cost priors added for --calibrate-obligations ($2.50) and --reconcile-obligations-deep ($1.50).

  • 2026-09-15 (run 2, after redeploy, $2.61) — still BELOW BAR, but the shape moved: recall 0.75 (3/4 abandoned setups opened; the violet seal is still the miss), yet all four resolved payoffs "not recognised" with the judge running, and control MODERATE+ rose to 205 (455 opened, 198 closed, 257 outstanding). Two more tail-cuts, same class as run 1: (3) the judge clamped every candidate passage to 600 words; the payoff beats are ~1,100 words with the payoff last. v2 lists a long candidate as consecutive numbered parts, gates quotes against the whole beat, and merges one beat's parts into one cache row (closes > advances > not_addressed); obl-judge-v2 voids the cut verdicts. (4) At scan time the extractor is shown the 40 most URGENT open rows; on a 257-outstanding book no book-end plant ever made the list, so the payoff beat could not "touch" the debt it pays. SelectForListing now keeps 25 urgent slots and fills 15 with rows whose content words occur in the beat (falling back to urgency); obl-extract-v3 restamps. Tests pin both. The 205 controls remain unmeasured until the judge sees whole beats. Run 3 (~$2.6) needs another redeploy.

  • 2026-09-16 (runs 3 and 4 — void; they measured nothing) — The harness never reset ledger state between runs, so every run inherited the verdicts of the one before it. Run 3's judge CLOSED the injected stopped-clock plant on "…smoked a cigar and waited behind a tree…" — a real sentence from the book, and entirely unrelated to the debt. Run 4 then scored that same row as a miss, on run 3's close event, with no run-4 event on the row at all. A false close is terminal: ObligationResurfacingJudge only ever revisits Open|Advanced rows, so nothing could correct it. Do not reason from any run 1–4 number; cumulative spend to that point was $14.38 for no measurement. Two fixes: ObligationCalibrationService.ResetLedgerAsync drops machine rows, their journal and the judge cache and clears ObligationScanHash before every score — preserving authored and author-locked rows, and refusing any book outside the gutenberg universe — and the judge gained a relevance veto. Grounding is not relevance: QuoteGrounding.Contains proves a quote is real text in the candidate, never that it answers the debt. SharesContent now vetoes a closes/advances verdict whose quote shares no content word with the obligation; obl-judge-v3 restamps so ungated v2 verdicts are re-judged rather than replayed. The underlying product risk is noted, not fixed: outside calibration, one bad auto-close still drops a real obligation forever.

  • 2026-09-16 (run 5, $3.10) — the first valid measurement. BELOW BAR: TP 4, FN 0, resolved mis-flagged 2, control MODERATE+ 103 (9.85/10k), P 0.037, R 1.000, F1 0.071. Ledger 229 rows (not the old cumulative 563): 134 outstanding, 95 closed. The stopped clock came back Advanced (flagged) and the violet seal Open (flagged) — the first time in five runs — but the extractor prompt was byte-identical to run 4's, so the credit belongs to the reset, not to any prompt fix. Diagnosis corrected: over-extraction is not the dominant term. Close rate by chapter is ch1 82% against 15–59% for every later chapter, across twelve structurally identical self-contained Holmes stories. Doyle did not get worse at resolving his own plots; the only variable is how many older debts compete for the same 40 slots, and outstanding-left-behind climbs with it (ch1 leaves 3, ch7 15, ch12 18). Reading the starved rows confirms they are real debts the story does pay — the masked visitor who unmasks as the King of Bohemia, Spaulding's cellar (he is digging the tunnel), the typewriter wear that convicts Hosmer Angel, the locked attic room. The extractor was right; the model was never shown the row it was paying. SelectForListing now fills three tiers — LOCAL (≤15 debts opened in this chapter, newest origin first), then lexically relevant rows from anywhere, then urgency for the remainder — on the principle that proximity, not urgency, predicts which debt a beat pays. obl-extract-v5 restamps. Left untouched so run 6 stays attributable: the judge's candidate finder, MaxOpenListed itself, and the off-page back-reference false-plant class ("the Tankerville Club scandal", "the Irene Adler photograph") — prose gesturing at stories it will never tell, which can never be paid and should never be opened.

  • 2026-09-16 (run 6, $3.03) — VOID: the extractor read 55 of 96 beats. The printed score was TP 2, FN 2, resolved mis-flagged 1, control MODERATE+ 99 (9.47/10k), P 0.020, R 0.500, F1 0.038 — and it describes 57% of a book. The trial balance immediately after the run reported ledger scan coverage : 55/96 beats; the calibration output itself said nothing, because it never reported coverage at all. ScanBeatAsync stamps a beat only when the extractor reports Evaluated — by design, so an outage re-scans rather than banking a bad read — but every Evaluated: false path in Parse returned silently (empty response, no JSON object, malformed JSON), and ExtractAsync discards all windows of a beat when any one of them fails. An unread beat was therefore indistinguishable from a beat that owed nothing. Leading suspect: truncation at maxTokens = 1400, since a response cut off mid-object has an opening brace and no closing one and lands on the "no JSON object" path — which would be the worst available sampling bias, because the beats that blow the ceiling are the ones opening the MOST obligations, so the instrument would be dropping the densest beats in the book rather than a random 43%. Do not attribute run 6's numbers to the locality fix, and do not compare them with run 5. Not yet known: whether run 5 (and every earlier run) had the same partial coverage — nothing recorded it, so runs 1–6 all lack a coverage figure. Fixed here, without guessing at the cause: Result.Failure records why a read failed and ExtractAsync logs a warning naming the window, beat length and reason; Score carries BeatsTotal/BeatsRead, CouldNotLook is true when they differ, and MeetsBar is false whenever CouldNotLook, so a partial run is void — not passing and not failing; the CLI prints beats READ by the extractor: 55/96 and RESULT: VOID — COULD NOT LOOK. prose --obligations coverage --slug <slug> reports read-only and free which beats were read and the length distribution of the read vs unread groups — the evidence that confirms or kills the truncation hypothesis from run 6's residue before run 7 spends anything.

  • 2026-09-16 (GCTOC, ~$0.15 — root cause of the unread beats, found by changing books) — Run 6's coverage hole had an unknown cause and a wrong leading suspect (truncation at the token ceiling). Building the ledger on GCTOC — A Tale of Two Cities, Book the First (17 beats, 17,300 words, ~1,018 words/beat — structurally comparable to GCSH's ~1,100, but one continuous narrative rather than twelve self-contained mysteries) named it in one command: scanned 17 of 17 … not evaluated 5, against exactly 5 beats over the 6,000-char window limit. Computing each split: 6,255 → 5,900 + 355; 6,126 → 5,985 + 141; 6,248 → 5,931 + 317; 6,277 → 5,883 + 394; 6,168 → 5,881 + 287. A beat a few percent over the limit splits into a real window plus a scrap; handed a scrap with nothing in it the model answers in prose rather than JSON, Parse returns Evaluated: false, and one unreadable window voids the whole beat — discarding the ~5,900 chars that read perfectly and leaving it unstamped. It is the opposite end of the beat from the suspected truncation. GCSH's beats measure 5,525–6,825 chars, so roughly half cross 6,000: that is exactly why run 6 read 55 of 96. The bias is systematic and was invisible — the beats that vanish are the long ones, in every book. Fixed: SplitIntoWindows folds a trailing window under max/5 into its predecessor (an oversized window costs a few hundred tokens; a scrap costs the beat), obl-extract-v6 so every over-length beat is re-read. Lesson: a second fixture of a different shape found in one cheap command what six paid runs on one book could not.

  • 2026-09-16 (harness defects fixed between runs 6 and 7, no LLM cost) — Two instrument problems that no amount of spending would have surfaced. (1) The scorer could not fail. Of the eight injections only the four ABANDONED ones were ever tallied TP/FN; of the four RESOLVED ones, only a row left outstanding scored (as a mis-flag), while both "never opened" and "opened and Closed" printed ok and fell through as neither. So recall was a statistic over half the manifest, and a planted debt the extractor missed entirely — the worst outcome available — counted as a PASS. An extractor that missed every resolved injection outright would have reported exactly run 5's "recall 1.000". Classify now scores all three resolved outcomes (never opened or left in a non-paying terminal state → FN, opened-and-Closed → TP, still outstanding → mis-flag as before), and is a pure static tested without the LLM pipeline — the reason the gap survived five paid runs is that nothing could exercise the scorer without spending $3. Score.ScorerVersion (obl-score-v2) is stamped on every result and printed: run 6 and everything before it are v1, and their recall is not comparable to run 7's. (2) prose --obligations candidates --id <guid> [--payoff-beat <guid> …] is a read-only retrieval diagnostic — no writes, no LLM call. "Payoff not recognised" hides two opposite defects: if the judge was shown the payoff text and still said not_addressed, the fix is recognition or granularity (a ~1,100-word beat carries one embedding, so a two-sentence payoff inside it is a rounding error against the whole beat's meaning); if the finder never surfaced the beat, the judge never had the chance, and the fix is retrieval budget. The verse-index proposal is only worth building under the first diagnosis. The command prints the candidate set, the ObligationJudgeCache rows behind it, and per payoff its lexical score and rank over every later beat (flagging when it sits below the finder's score >= 2 floor) plus its rank in a deep top-400 embedding sweep, so a payoff outside the k=8 budget reports how far outside rather than merely absent.

  • 2026-09-16 (the retrieval diagnostic, run at last — free, no LLM, and it answers the question the wrong way round) — --obligations candidates had never been run. §8's own note said "the verse-index proposal is only worth building under the first diagnosis" — i.e. only if the judge was shown the payoff and still failed. It was shown nothing. Both GCSH resolved injections left outstanding come back NEVER RETRIEVED, for two different reasons, and neither is recognition:

    1. Brass whistle (01a0ab6e3c1a7db7bebaa99ac13ca174, Open, origin Ch1 pos 2). Payoff is Ch11 pos 81. Lexical rank #15 of 93 — outside CandidatesPerObligation = 8. Embedding: outside the top 400 of the corpus. The reason is not granularity — the index is empty. The first reading of this run blamed one-vector-per-1,100-word-beat being too coarse to locate a two-sentence payoff. That was wrong, and the diagnostic's own output disproves it: the deep sweep is FindSimilarBeatNodesAsync(…, k = 400) and it returned 12 hits — the same 12 in both probes. The gutenberg universe holds 116 beats (GCSH 96 + GCTOC 17 + GCNEG 3), and only 12 carry a row in ProseEmbeddings: precisely GCSH's twelve original one-per-chapter beats, from before it was split to 96. split_beat never backfilled the embedding index. Every candidate that scored (01a0514d7…) is one of those twelve chapter blobs; every split beat shows emb —. That also explains the flat 0.300–0.355 spread: it was not ranking passages, it was returning all twelve chapters in near-arbitrary order. The embedding tier of the retrieval finder has been operating on a 12-row index of chapter-sized blobs for this entire programme, and the payoff beat has no vector at all — it cannot rank because it does not exist to the query. Consequence: we still do not know whether beat-granularity embeddings are adequate, because they have never once been tested. The fix is a backfill, not an architecture.
    2. One-eyed cat (01a0ab7ba3ef7743b645f7c22197fbae, Open) — worse, and a cascade from the coverage bug that survives the coverage fix. Its origin quote is the payoff sentence: the row is anchored to Ch9 pos 64, the payoff beat itself. The finder only ever looks after the origin, so this debt is structurally unclosable forever. Traced end to end: the setup beat is 01a0a698e81f7935b8a05270b53b9e0a, Ch2 pos 12, 6,186 chars — over the 6,000 window, and on the unread list. So: the setup was never read → the debt was never opened → the payoff beat was read with no open row to close → the extractor opened a new row on the payoff → the judge can never reach behind it. The scorer then reported "payoff not recognised", blaming recognition for a coverage failure three steps upstream. Consequences. (a) The failure is retrieval, not recognition — in neither case did the judge ever see the text. But the first-order fix is backfilling the embedding index and making beat-creation maintain it, not building a finer one; the verse-index proposal stays unbought until an actually-populated index has been measured and found wanting. Raising CandidatesPerObligation and protecting the lexical tier from being crowded out are cheap and independently justified — the brass-whistle payoff placed #15 of 93 lexically, so lexical alone would have found it at k=16. (b) A new defect class: an obligation anchored to the wrong end of its own arc, which no amount of judge improvement can fix. Inside calibration ResetLedgerAsync clears it; on a real book there is no reset, so this corruption is permanent — the same shape of risk already logged for a false close. (c) Every "payoff not recognised" number in runs 1–6 is now suspect as mis-attributed coverage damage. (d) A silent-index class joins the silent-failure classes: a retrieval tier that returns 12 rows for a 400-row request reported no error, and nothing in six paid runs noticed. ProseEmbeddings coverage belongs in --obligations coverage alongside beats read.

  • 2026-09-16 (the backfill, and the answer: beat-granularity embeddings are ADEQUATE) — Hub redeployed (PID 54284) with obl-extract-v8, ObligationScanAttempts, Node.StructurallyComplete and prose --reembed --beats. Backfilling the index took seconds and cents: GCSH 96 beats, GCTOC 17. The corpus sweep went from 12 hits to 113. Re-running the identical brass-whistle probe against the populated index:

    before after
    corpus hits (k=400) 12 113 (93 after the origin)
    payoff embedding rank outside top 400 #1 of 93 (similarity 0.415)
    payoff lexical rank #15 of 93 #15 of 93 (unchanged)
    in candidate set no YES
    verdict NEVER RETRIEVED RETRIEVED, NOT YET JUDGED

    One vector per ~1,100-word beat put the true payoff FIRST out of 93 later beats. The similarity spread is now a real gradient (0.415 at #1 → 0.297 at #64) where before it was a flat 0.300–0.355 across twelve chapter blobs — which is what a meaningless ranking looks like, and should have been the tell. So the verse-index / sub-beat-passage proposal is not needed, and that is now measured rather than assumed. Note also that the payoff's lexical rank did not move: at k = 8 lexical alone would still have missed it. The embedding tier found it — once it existed. The earlier note that "lexical alone would have found it at k=16" holds only at k≥16. The one-eyed cat is unchanged — BEFORE THE ORIGIN — against the same populated index, confirming the two defects are independent: a full index cannot fix a row anchored to its own payoff. Wrong-end anchoring needs its own re-anchor path. Remaining before the brass whistle can actually close: the judge has no cache row for that candidate at any prompt version, so a --deep pass must now be bought to find out whether it recognises the payoff it can finally see.

  • 2026-09-16 (GCTOC beat 3: the blind spot explained, and partly closed — ~$0.02) — Re-scanning the one beat under obl-extract-v8: opened 1, advanced 2, ungrounded discarded 3, not evaluated 0. It had reported opened 0 under v7. The beat was never empty — its entire harvest was being thrown out by the quote gate, and that reported identically to a beat that owed nothing. Defect confirmed; it is now an ObligationScanAttempt row with outcome read-all-discarded rather than an invisible one. Graded honestly against the §6a key, this is partial credit, not a fix:

    • The new row 01a0ad7351377d79ad2bd12ebfd16c65 carries key row #1's exact origin quote, "Jerry, say that my answer was, RECALLED TO LIFE." — but describes the debt as Jerry's identity, not what "RECALLED TO LIFE" means. Right quote, arguably the wrong debt.
    • Key row #2, Wait at Dover for Mam'selle, is still not opened at all.
    • 3 items were still discarded ungrounded. Quote fidelity on Dickens — em-dashes, archaic spelling, nested quotation — remains a live defect, now measured instead of hidden. A9's typed filter is not the fix for this; the quote gate needs a normalisation pass of its own. Do not report GCTOC as 14 or 15 of 15 on this evidence: one beat was re-read under v8 while the other sixteen still carry v7 rows, so the book is a mixture and not a measurement. PromptVersion is folded into ObligationScanHash, so every GCTOC beat is already due a re-read; a clean grade costs one full run (~$0.41).

  • 2026-09-17 (literature pass — seven research expeditions; four rulings changed on evidence) — The author asked for the field surveyed and the design derived scientifically rather than by taste. Condensed here; full version with citations in the plan file.

    (a) Our quote-gate discard rate is normal, and the specific cause is probably typography. Published baseline: 22–28% of an unconstrained model's "verbatim" quotes are not verbatim (ReClaim, NAACL 2025 Findings — CR 75.5/72.1/77.5). Every such measurement is on Wikipedia-style text. QuoteGrounding.Normalize collapsed whitespace but not Unicode, and Dickens is dense in em-dashes and curly quotes — so a substantively correct claim was mechanically invisible, which is exactly beat 3's opened = 0. Fixed this session (FoldTypography + TryAlign). The recovery strategy is the measured best: FullCite scored post-hoc word-Jaccard alignment 61.87 against 12.80 for prompting and 55.11 for constrained decoding — and constrained decoding refused on up to 56.6% of cases. CiteFix separately measured deterministic matching beating an LLM re-matcher 8× on quality at 100× lower latency.

    (b) RULING REVERSED — finer granularity is measurably WORSE on narrative. Decomposing beats into sentence-level "verses" is rejected for retrieval and decomposition alike. On GutenQA (100 Project Gutenberg narrative books), proposition-level chunking has the worst Recall@20 of five methods — 55.54, below plain paragraph at 64.34 (LumberChunker, EMNLP 2024 Findings), attributed to pronoun ambiguity in narrative. Citation granularity peaks at paragraph scale, 4–8 sentences, and fine granularity penalises larger models hardest (Llama-70B +70.1% from k=2→k=4; 120B +96.7%). Atomic units verify worse exactly where it matters: 22.4% accuracy on the not-supported class vs 38.8% molecular. The target unit is paragraph-scale. Our ~400-word beats are near it; the 6,000-char outliers are the defect.

    (c) THE <1% BAR HAS A MEASURED CEILING. FABLES (COLM 2024) is the closest published measurement of our exact task — claims about book-length fiction, annotators who read the whole book. Best auto-verifier, Claude-3-Opus with the entire book in context: F1 0.967 on faithful claims but 0.476–0.582 on UNFAITHFUL ones; GPT-4-Turbo 0.109. Authors' verdict: "too poorly to be a reliable auto-rater." The irreducible part — annotator agreement on subjective narrative claims is 26.56% (Subbiah et al. 2025). <1% is unreachable on the interpretive fraction by anyone, by any method. It must be scoped to the objective, span-checkable fraction, and the two reported separately, or the noise floor silently caps every number we publish.

    (d) Chekhov's gun is misquoted, and the real rule fixes our precision problem. The 1889 letter criticises a superfluous monologue, not a prop, and its operative clause is «Нельзя обещать» — "one must not make promises." The act-numbered version comes from two memoirs published 15 and 22 years later; Chekhov's own plays leave guns unfired on purpose. The defensible rule is about SALIENCE, not inventory: a text must not foreground what it will not use. That is the fix for our texture false-plants — the guard's blunderbuss was never foregrounded, so it was never a debt. Corroborated computationally: filtering non-salient events improved narrative prediction by ~5 points, with a measurable over-filtering point (Zhang et al., EMNLP 2021).

    (e) Do not build narrative event chains. The Chambers & Jurafsky lineage fails on stories: a unigram model beats PMI chains on narrative cloze (rank 400 vs 1115); on ROCStories C&J chains score 0.478, below the 0.513 "always pick the first option" constant; on non-news narrative both published script models lose to a most-frequent baseline. What finally beat it was typing the participants, not refining the event representation. Also: a verbs-only literary event baseline gets recall 76.2 at precision 17.7, and "thought" is a realis event only 32% of the time. A lemma is not an event.

    (f) Coreference is the hard ceiling, and it dictates an architecture. Whole-book coreference tops out at 67.0 CoNLL F1 fine-tuned and 41–47 off the shelf, while the same models on 1,500-token windows score 82.2 (BookCoref, ACL 2025) — the entire 15–21 point loss is in stitching. So: window, then merge globally. Two further consequences: MUC is worthless at book length (BookNLP scores MUC 83.1 with CEAF_φ4 2.4 — a catastrophic merge reading as a good score, the same failure shape as "zero findings can mean could not look"), and a bounded memory of ~20 entity slots suffices, costing ~1.4 F1, because LitBank documents never have more than 18 entities simultaneously active. Singleton warning: OntoNotes annotates 0% singletons by design — and a referent introduced once and never returned to is precisely what we hunt. The curtain girl is a singleton.

    (g) The entity "web" is sparser than the metaphor, and professional practice is not a graph. Fictional network density collapses with cast size (plays 0.21–0.72; epics 0.01–0.06), fictional networks "do not exhibit realistic topological properties", and Moretti never published node/edge counts while his own epilogue calls the weighted-directed version "a dead end, pure and simple." Every professional continuity system — the Star Wars Holocron (FileMaker, ~120k entries), OHOTMU, Dragonsteel's Cosmere timeline — is a flat record store with a small field set plus free prose. The one to steal is Karen Ahlstrom's row schema, which is our obligation row independently invented: every row carries the quote that justifies it, the chapter that quote is in, and a flag for whether the event happened then or is merely referenced. And every one of those systems degrades into "ask the one person" — DC's continuity index was 3×5 cards, and someone threw the drawer out during an office move.

    (h) Build nothing on percentage marks. TRIPOD annotated 99 films against Hauge's five turning points: a content-blind baseline placing beats at the observed mean beats the textbook percentages outright, and is competitive with a neural model that reads the screenplay. Two of five turning points are significantly displaced; the middle three carry SDs of ±8 to ±12 points. The schemes agree that 50% is the middle and assign it five incompatible functions. The three craft rules that are checkable: Swain/Bickham stimulus→reaction clause ordering (nearly pure dependency-parse), McKee's per-scene value-charge turn (he supplies the algorithm and names exposition as the dominant false positive), and Sanderson's First Law read as plant-before-payoff, which maps straight onto PlantPayoffs.

  • 2026-09-17 (GCNEG run for the first time — ~$0.06 — and §6 names the wrong fixture) — The negative control had never been scanned in the programme's history (0/3). Rescanned: scanned 3 of 3 — opened 9, advanced 5, closed 2, ungrounded discarded 0, not evaluated 0.

    Nine rows from a "negative control" looks catastrophic and is not. Reading the fixture before judging it — the discipline that has paid every time this week — GCNEG's own header says what it is: "Synthetic negative-control fixture for the Full-Order Read instrument (docs/LOGIC.md §10 Phase 2): Ch1/Ch3 are deliberately engaging, Ch2 is deliberately flat." Its design axis is engaging-vs-flat, not owes-debts-vs-owes-nothing. Ch1 is a locked-room murder saturated with genuine promises ("tell her it's tonight", the untouched key, the unburned scrap, "someone was still lying"); Ch3 solves the locked room and opens two more. A competent ledger must open rows there.

    Where the rows landed is the actual result: 7 from Ch1, 2 from Ch3, and ZERO from Ch2. Ch2 — the deliberately flat Margaret beat, and at 1,903 chars the longest of the three — produced nothing. Every Ch1/Ch3 row is defensible on inspection ("Who is 'her' and what event is planned for 'tonight'?"; "How was Whitfield killed in a locked room?"), two are correctly Closed — Ch3's fishing-line mechanism paying Ch1's locked-room question — and there are no texture false-plants at all: no row on the butler's wringing hands, the cold ash, or the painted window as scenery. Worth noting against §6a's Dickens result, where texture false-plants were ~7 of 27: GCNEG is modern plain prose, and this is consistent with the salience finding above — Victorian prose foregrounds texture far more, so the over-extraction is partly a property of the register.

    Two findings:

    1. The extractor passes a true obligation negative control cleanly, and demonstrates a full open→close arc inside one fixture — the first time the programme has shown either. Caveat: this is one beat, n=1; it is a datapoint, not a validated false-positive floor.
    2. §6 is wrong: it names GCNEG as a calibration fixture for this instrument, and GCNEG belongs to a different one. Only 1 of its 3 beats is a valid obligation negative control. The programme has no purpose-built obligation negative control, and has been listing a borrowed fixture with an orthogonal design axis as though it were one. A real one is cheap to build: competent, readable prose that closes everything it opens in-scene and foregrounds nothing it does not use — Ch2 is the template, and it needs to be ~20 beats, not one.

  • 2026-09-17 (GCOBN built and scanned — the purpose-built obligation negative control §6 called for — and it is the cleanest, worst result yet) — Built gutenberg-calibration-obligation-negative-control-synthetic-01a0af36 (NodeCode GCOBN): 20 beats, one ordinary day for a retired man, every beat hand-written to close whatever it raises in-scene ("nothing decided and nothing owed on either end", "that was the whole of their acquaintance", "the ordinary amounts he expected") and to introduce no object with more weight than the sentence describing it. Rescanned (~$0.2): scanned 20 of 20 — opened 25, advanced 32, closed 1, ungrounded discarded 0.

    25 rows from 20 beats engineered to owe nothing. Reading every row (the discipline that has paid every time this week): all 25 are the same failure shape. The extractor is not finding plot debts — it is converting any described object or explicit closure sentence into an interpretive "why does this matter" question. It opened a row asking why the room's temperature "signifies" neutrality, why the bench "will come to represent" something because it is empty, why Walter's never-asked name for the register clerk "signifies about Walter's life" — and, most tellingly, it opened rows on the closures themselves: "nothing decided and nothing owed on either end" became "whether this non-decision will have consequences"; "that was the whole of their acquaintance" became "whether she will return or represent a crack in his isolation." The one row it Closed is itself the same artifact wearing the other hat: it opened and closed a row on "nothing about it was different from any other Tuesday's" — treating a sentence that denies anything happened as both a promise and its own payoff.

    This is not a recall problem and no amount of explicit in-text closure defeats it. Every sentence in GCOBN whose entire grammatical job was to foreclose a question was read as proof one was open. The extractor has no operating notion of salience (RFC 0013 §8g's Chekhov correction — "do not make promises," not "do not describe objects") — it is scoring textual attention itself as the plant. On this fixture the false-positive rate is effectively 24/25 (96%), an order of magnitude past the GCTOC §6a Dickens texture-false-plant rate (~7/27, ~26%) and two past the <1% bar. GCOBN is now the calibration fixture for this number — GCNEG's borrowed 1/3 valid beat is retired from this role.